a simulation study of tcp performance over umts dl
TRANSCRIPT
A Simulation Study of TCP Performance overUMTS Downlink
B. J. Prabhu�, E. Altman
�, K. Avrachenkov
�, J. Abadia Dominguez
��INRIA, 2004 Route des Lucioles, 06902 Sophia Antipolis Cedex, France.
Email: � bprabhu, altman, k.avrachenkov � @sophia.inria.fr�France Telecom R&D, 38, rue du General-Leclerc, 92794 Issy-les-Moulineaux Cedex 9, France.
Email: [email protected]
Abstract— Data transfer on the UMTS downlink can be doneeither through dedicated channels (DCH) or shared channel(FACH). The data transfers on the DCHs can receive very highthroughputs. However, a setup time of the order of 250ms isrequired before data transfer can begin on the DCH. For veryshort transfers it may be better to use the shared channel.However for long data sessions it might be better to allocatea dedicated channel. In this paper, we propose a threshold policyto determine which sessions should use the dedicated channels.Initially, we use the FACH for a given connection. Then if we getan indication that the current burst might be long (for example,we observe a long queue from that source), then beyond somethreshold we shall try to allocate a DCH to that connection(if there is one available). We also present and evaluate theperformance of algorithms in which a timer is set when the queuesize falls below a threshold and the connection is on DCH. Theuse of timer allows a connection to remain for a longer durationon DCH and thus improve performance. We use ns-2 simulationsto obtain the performance measures for the threshold and timerbased policies.
I. INTRODUCTION
UMTS is the ����� generation of cellular wireless networkswhich aims to provide high speed data access along withreal time voice calls [1]. On the UMTS downlink, data canbe transferred through dedicated channels (DCH) or a sharedchannel (FACH). Dedicated channels offer higher transferspeeds but require a setup time which is significant (of theorder of 250ms). Shared channels, on the other hand, havea low setup time and also low transfer speeds. For sporadicpackets (i.e., files or transfers that are very short) it is efficientto use the shared channel. However, for long data sessionsit may be better to allocate a dedicated channel. Most datatransfers over the Internet are performed today with TCP. Dataconnections on the Internet are known to be bursty where thebasic burst size is best described with a Pareto distribution:� � ����������������������! #"%$'&(��)+*-,
where$
is the minimum size and .0/214365 (for Internettraffic) [2]. We use the traffic description of Appendix B of [3]which is the standard for UMTS data structure. It is also basedon the Pareto distribution for the transfer size. Our startingpoint is the observation that long bursts are in some sense”predictable”: if we observe that the current burst has alreadytransfered many packets, say 7 , then the distribution of the
remaining size of the burst 8 9��������:�������:; 7 is given by�<" 8 � 7>= �@? 8 � 7 )A CB 77>= �ED *so for a fixed value of
�, as 7 increases (to infinity),
�@" 8 �7F= �<? 8 � 7 ) increases to 1. (This is a general propertyof heavy tailed distributions [4]). Therefore, it seems naturalto implement the following policy: start using the FACH fora given connection if a DCH is not available. Then if weget an indication that the current burst might be long (forexample, we observe a long queue from that source), thenbeyond some threshold we shall try to allocate a DCH tothat connection (if there is one available). When the queuesize of a connection falls below a threshold we switch itback to FACH. We also present and evaluate the performanceof a modified threshold algorithm. In the modified thresholdalgorithm, instead of switching immediately when the queuesize falls below a threshold, a timer is set and the connectionremains on the DCH for this period. If there are no new arrivalswithin this timeout period, the connection is switched back toFACH. The timer is used in order to allow the ACKs to reachthe TCP source and new packets to be released by the source.
The rest of the paper is organized as follows. In Section II,we present the network configuration used in the simulationsand describe the methods used in the implementation of thecode. In Section III, we describe the source traffic model,the simulation parameters and give the values used in thesimulations. In Section IV, we present the results of thethreshold algorithm. In Section V, we describe the modifiedthreshold policies wherein a timer is used along with thethreshold on the DCH. In Section VI, we present the results ofthe modified threshold policies and compare them with theresults of the original threshold policy. In Section VII, wepresent the conclusions obtained from this work.
II. NETWORK MODEL
We consider a network model with GIHKJ+L TCP sources whichneed to send data to mobile receivers. We assume a singlecell scenario with one base station and several mobile stationswhich act as destinations for TCP traffic. The TCP sources areassumed to be connected to the base station of the cell with ahigh speed (5mbps, 30ms) link. The base station can transferdata from a TCP source on either DCH or FACH at a given
time. There is one FACH and G � J�� DCHs in the system. TheFACH is a time division multiplexed channel. In addition toany TCP connections which may be present on a FACH, thereis signaling traffic which must be transmitted on the FACH.The signaling traffic has priority over the TCP connections.During the silence periods of the signaling traffic, data fromone or more TCP connections can be transmitted on the FACH.On the FACH, data from the TCP connections is assumed tobe transmitted in a round- robin fashion. If all the DCHs havea TCP connection configured, a connection on DCH shouldbe first switched to FACH before a session from FACH canbe switched on to a particular DCH. This means that a switchcan take up to 500ms (if there is already a TCP connectionconfigured on the DCH).
Switching from one channel to another is costly in time andin signaling. In the model we assume that there exists a queuecorresponding to each TCP connection in the base station. Thebase station is hence able to track the queue length of eachconnection. During the switching time (of around 250ms) fromone channel to another, no packets from the queue of the TCPconnection being switched can be transmitted. We, therefore,try to avoid quick switching and propose the following channelallocation policy. There are two thresholds
� � and���
. If thenumber of packets in the queue of a connection using FACHreaches
� � and a DCH is available then we initiate a switch tothe DCH. If the number of packets in the queue of a connectionusing DCH drops below
� �and the number of allocated DCHs
equals the G � J�� , then we switch the connection back to FACH.The simulation setup for the network is presented in Fig. 1.
384
kbps
, 10m
s
384
kbps
, 10m
s
384
kbps
, 10m
s
33 k
bps,
10m
s
5 m
b, 3
0ms
5 m
b, 3
0ms
5 m
b, 3
0ms
Mobile receivers
Core Network
Denotes virtual link Denotes virtual node
1SWT SWT2 SWTN
TCP TCP TCP1 2 N
FACH IN
FACHOUT
DST1 DST DST2 N
CBRSRC
CBRDST
Fig. 1. Simulation Setup
Each TCP source node,������
is connected to a routing nodecalled Switch ( � ���
). � ���is present inside the base station
and can be connected either to the �� �������or directly to
the TCP destination via the DCH. The ��� �node is created
to simplify the simulations and may not be present inside a
real base station. The �� ��� ���is another virtual node which
simulates the prioritized round robin service discipline takingplace on the FACH. This is a service discipline which we had toimplement. In this discipline, the node �� ��� ���
gives priorityto the traffic from CBRSRC while serving the packets fromthe ��� �
’s (only those which are currently not transmittingon DCH) in a round-robin manner. We note that there areno queues at �� �������
and all the packets are either queuedat � ���
or at the��� 8�A8 �
. The��� 8�A8 �
simulates aconstant bit rate source of signaling/control traffic. It generatespackets at rate 8�� � � and is assumed to be present within thebase station. The signaling traffic flows from the base stationto the mobile receivers. Even though we model the destinationof the signaling traffic (
��� 8"!# �) as one node different
from !$ � �, we note that it does not affect the simulations as
simultaneous transfer of data and control packets to the samemobile receiver is possible when different channels are used.The links ��� � ; �� ��� ���
and��� 8�A8 � ; �� ��� ���
are virtual links within the base station and thus have zerodelay. We note that the data from � � �
to !$ � �can take two
different routes i.e., ��� ��; �� ������� ; �� ����%�&(' ; !$ ���(via FACH) or � ���!; !# ���
(via DCH). At any given timeonly one route from the above two can be active. Although inthe simulation scenario we have as many DCH links as TCPsource nodes, the simulation allows us to connect not morethan G � J�� DCH channels at a time, which may be chosenstrictly smaller than the number of TCP sources ( G HKJ+L ). In thesimulations we switch over from FACH to DCH by changingthe cost of the links and recomputing the routes. This is doneas follows: Initially, the cost of the direct path from the Switchto the TCP destination is set to 10 and the cost of all other linksto 1. Hence, the traffic gets routed through the FACH. Whena switch is desired, the cost of the DCH is set to 1 and theroutes are recomputed. This activates the DCH and the trafficgets routed on the DCH.
III. INPUT FOR THE SIMULATION
We use ns-2 [5] for our simulation study with the followingparameters:
) Number of available DCH channels ( G � J�� ) is taken to be1. The number of simultaneous TCP connections, G>HKJ+Lvaries between 2 to 8.) Values of threshold
� � for switching between channels isvaried between 0 and 15. A threshold value of 0 indicatesthat we try to switch a connection to DCH as soon as itarrives.) The duration of the simulation is taken 200000 secs inorder to reach stationarity. The large time needed to getgood accuracy is partially due to the nature of the longrange dependence of the traffic.) The time it takes to switch between channels ( !*�,+ ) is250ms.) We consider the background (non TCP traffic source)traffic source, that uses the FACH, to be CBR source withrate 8 � � � 5.- kbps. It sends a 1kB packet at an intervalof . & � secs. It has non pre-emptive priority over TCP.
TABLE I
NETWORK PARAMETERS
�������1�������
384 kbps�������33 kbps� ����24 kbps���
1� ��.25 s
TABLE II
TRAFFIC SOURCE PARAMETERS
����� ��� � �� � ! " �# �
0.33$&% �(' �30 kB)
1.1# )+* ��-,�.320 B
) The TCP connection traffic model is as follows: On a TCPconnection, data arrives in bursts. The number of packetsin a burst has a Pareto distribution. The shape parameter isgenerally taken to be between 1.04 and 1.14 [2], [6]. Wetake the shape parameter to be
$ .�/ . [3] and the averagefile (burst) size is taken to be 10(2 � ��3 $ �
[2]. We donot take all the parameters from [3] as they are not adaptedto TCP traffic. The model in [3] considers packets of size81.5 bytes whereas TCP does not send short packets [8].A TCP connection alternates between ”ON” and ”OFF”states. In the ON state, the interarrival time betweensuccessive bursts is exponentially distributed with mean�14�5 � . At the end of each burst, the connection goes intoOFF state with probability
&6�787. It remains in the OFF
state for an exponentially distributed duration with mean� � � � . At the end of this duration the connection goes toON state which is marked by the arrival of a burst. Fig.2 shows the traffic model used at each source node.) The TCP packet size is taken to be 320 bytes which is ofthe same size as a RLC data segment in UMTS [9].) The schedule of service of packet in the FACH is takento be round robin. The implementation in ns-2 is notstandard and we had to change some link element featuresof ns-2 in order to implement this schedule.
Tables I and II give the values of the network parameters andtraffic parameters, respectively.
IV. SIMULATION RESULTS I
Fig. 3 shows the average burst delay versus� � for different
number of TCP connections, G HKJ+L . For a given value of G HKJ+L ,it is observed that initially the delay reduces and then increaseswith increasing values of
� � . This suggests an optimal valueof the threshold at which the delay is minimum. At highervalues of
� � an increase in the burst delay is observed becausea higher value of
� � implies more time is spent in theFACH. The FACH is a low bandwidth channel which has
OFFON
(1 − P Poff*1/Ton)*1off /Ton
1/Toff
Fig. 2. Traffic source model
0
2
4
6
8
10
12
14
16
18
20
22
0 2 4 6 8 10 12 14 16A
vera
ge b
urst
del
ay (
s)
Threshold (Th)
Ntcp = 2Ntcp = 3Ntcp = 5Ntcp = 8
Fig. 3. Average Delay vs.�9�
.�����:�<;>=
,� � ;@? " ��AB�
.
high priority signaling traffic on it. This results in low averagebandwidth being shared amongst the TCP connections. For aTCP connection, the switch to DCH is based on its the currentbuffer size which in turn depends on its current window size.The current window size is incremented whenever an ACKis received by the sender. When a TCP connection is on alow bandwidth link, the window builds up slowly due to delayin receiving an ACK. This slow buildup of the window sizeresults in slow buildup of the current buffer size. As the valueof
� � is increased, a TCP connection has to spend more timeon the slow FACH before it builds up its buffer size resultingin a higher delay.
We also observe a decrease in the burst delay for smallvalues of
� � . This dip is observed due to the high speed ofthe DCH and the delay before ACKs can reach the source.Before the source can release more packets to the DCH queueit must receive an ACK from the
���� �(C �%H . However, there is alink delay and transmission delay before the ACK is receivedby the source. If, in this period, the DCH queue becomesempty, the session is switched back to the FACH. As soon asthe source receives the ACK it releases packets to the queuethus prompting a switch back to DCH. This redundant switchconsumes D�3�3�E � and thus adds to the delay of the burst.However, if there are enough packets present in the queue suchthat the queue does not become empty prior to the arrival ofthe ACK, there is no switch back and forth. The number ofpackets which ensures that this back and forth switch does nottake place is 5 and hence we see an optimal value at
� � - .
0
2
4
6
8
10
12
14
16
0 2 4 6 8 10 12 14 16
Thr
ough
put (
kBps
)
Threshold (Th)
Ntcp = 2Ntcp = 3Ntcp = 5Ntcp = 8
Fig. 4. Throughput vs.�9�
.�����:�<; =
,� � ; ? " ��A �
.
For a given value of� � , the average burst delay increases
as a function of G HKJ+L . As the number of TCP connections onthe FACH increase, the available bit-rate for each connectionreduces thereby increasing the time taken to increase thewindow size and hence to exceed the threshold.
Fig. 4 shows the throughput as a function of the� � for
different values of the number of TCP connections, G HKJ+L . Thethroughput first increases and after an optimal value of
� �it decreases. The improvement in throughput at the optimumthreshold is more prominent compared to improvement ofdelay. For a given threshold, the throughput decreases withincreasing number of TCP connections.
V. MODIFIED THRESHOLD POLICY
In the previous section we observed that the use of theoriginal threshold policy resulted in a switch from DCH toFACH even though the source had packets to send. Thisresulted in a poor delay performance for the policy. In thissection we propose a modified threshold policy which aimsto ameliorate this flaw of the original threshold policy. In themodified threshold policy, the switch from the FACH to DCHtakes place according to the same criterion as the originalthreshold policy. However, when the queue length at ��� �drops below
���and the connection
�is on DCH we start a timer
for a duration� 6�5 H . If there are packet arrivals at the queue
( ��� �) during this period, we reset the timer to 0. We simulate
two policies which decide on the switch back to FACH.In the pre-emptive scheme, if during the timeout period there
is another TCP connection which requires a switch to DCH,we initiate a switch back to FACH.
In the non pre-emptive scheme, we wait till the end of thetimer before we initiate the switch back to FACH. Also, at theend of the timer if there are no other requests for the DCH,we start the timer for a duration of
� 6�5 H thus allowing theconnection to remain on DCH.
A timeout value of 3 results in the original threshold policy.The use of timeout periods is motivated by the fact that there isa delay before the ACKs reach the source. Hence, even thoughthe queue at the base station is empty the source may have
2
4
6
8
10
12
14
0 2 4 6 8 10 12 14 16
Ave
rage
del
ay (
s)
Threshold (Th)
Ndch = 1, Ntcp = 5
Dsw = 0.25s, To = 0.0s, peDsw = 0.25s, To = 0.1s, pe
Dsw = 0.25s, To = 0.1s, npe
Fig. 5. Average Delay vs.�9�
.�����:�<;>=
,� � ��� ; "
,� � ;@? " ��AB�
.
6
8
10
12
14
16
18
20
22
0 2 4 6 8 10 12 14 16
Ave
rage
del
ay (
s)
Threshold (Th)
Ndch = 1, Ntcp = 8
Dsw = 0.25s, To = 0.0s, peDsw = 0.25s, To = 0.1s, pe
Dsw = 0.25s, To = 0.1s, npe
Fig. 6. Average Delay vs.� �
.� ���:� ;>=
,� � ��� ;�� , � � ;@? " ��AB�
.
packets to send and may be waiting to receive the ACKs beforeit sends the data packets. The timer, to some extent, waits forthe ACKs to reach the source and the source to release thepackets.
VI. SIMULATION RESULTS II
Figs 5 and 6 show the average delay performance of thepre-emptive (pe) and non pre-emptive (npe) schemes of themodified threshold policy. We again note that the case of
� 6 3 corresponds to the original threshold policy. From the figures,we observe that the modified threshold policies also have aoptimal value of the threshold as was observed in the caseof the original threshold policy. As compared to the originalpolicy, the modified policy with pre-emptive scheme performsmarginally better. As the number of TCP connections increases,the probability of a request for a switch also increases. Hence,the pre-emptive scheme, in which a switch is initiated as soonas there is a request, performs very close to the original policywhen G HKJ+L �� . With G HKJ+L D , the duration of the idle periodon the DCH (i.e., when the timer is on and no requests havebeen received) is longer and thus the connection is able toreceive packets from the source and remain on the DCH andavoid a costly (in terms of time) switch back to FACH. The
5
6
7
8
9
10
11
12
13
14
0 2 4 6 8 10 12 14 16
Ave
rage
bur
st d
elay
(s)
Threshold (Th)
To=0.0, peTo=0.1, pe
To=0.1, npe
Fig. 7. Average Delay vs.�9�
.�������<; =
,� � � � ; "
,� � ; " �(��A �
.
8
10
12
14
16
18
20
22
0 2 4 6 8 10 12 14 16
Ave
rage
bur
st d
elay
(s)
Threshold (Th)
To=0.0, peTo=0.1, pe
To=0.1, npe
Fig. 8. Average Delay vs.� �
.� ����� ; =
,� � � � ; � , � � ; " �(��A �
.
non pre-emptive scheme performs better than both the originalpolicy and the pre-emptive scheme. The performance gains areclose to 5�3 �
at� � 3 and G HKJ+L D . In the non-preemptive
scheme, the base station waits till the end of the timer before itinitiates a switch. The timeout duration ensures to some extentthat the TCP session has ended and thus can be switched backto FACH. Hence, a TCP session which is once switched toDCH remains on the high speed DCH till the end of the sessionthereby reducing its average delay.
Figs. 7 and 8 show the performance of the policies when theswitching delay, ! + is D�3 3�E � . At a higher switching delay,the optimal threshold shifts to right. As the cost of shifting toDCH and back is high, it is preferable to remain on the FACHwhen the connections have fewer packets to send. However,we still note that the non pre-emptive scheme performs betterthan the other schemes. With a higher switching delay, theperformance gain obtained by using the optimal threshold ishigher and is observed to be around .+3 �
compared to the nothreshold case.
Figure 9 shows the average delay with G � J�� 5 , G HKJ+L �� ,and ! + 5�D�3�E � . We observe a similar behavior as in thecase of G � J�� . .
2
4
6
8
10
12
14
16
18
20
22
0 2 4 6 8 10 12 14 16
Ave
rage
del
ay (
s)
Threshold (Th)
To = 0.0To = 0.1, pe
To = 0.1, npe
Fig. 9. Average Delay vs.�9�
.�&�����<; ?
,� � ��� ; � , � � ; ? " ��AB�
.
Also, in general, we observe that there is significant loss indelay performance at threshold values higher than the optimalvalue.
VII. CONCLUSIONS
In this paper, we evaluated, through simulations, the per-formance TCP data transfers on the UMTS downlink. Weused a threshold policy to determine which TCP connectionsshould use the high speed dedicated channels. The resultsindicated the presence of a optimal value of the threshold forminimum burst delay. However, the threshold policy resulted inredundant switches between the FACH and DCH, and thereforein increased average delay for the file transfers. We used a timeron the DCH to prevent these frequent switches. We observedthat the use of a timer on the DCH resulted in improved delayperformance for the bursts (files). The results indicate thatfor the simulated parameters the use of a threshold did notsignificantly improve the delay performance and suggest thatuse of a timer with a threshold can give improved performance.
REFERENCES
[1] H. Holma and A. Toskala, WCDMA for UMTS: Radio Access for ThirdGeneration Mobile Communications, John Wiley & Sons, 2001.
[2] B. A. Mah, ”An Empirical Model of HTTP Network Traffic”, Proc. IEEEINFOCOM’97, April 1997.
[3] UMTS 30.03 version 3.2.0, Technical Report TR 101 112, ETSI, April1998.
[4] C. M. Goldie and C. Klppelberg, “Subexponential Distributions”, in APractical Guide to Heavy Tails: Statistical Techniques and Applications,ed. R. L. Adler, R. Feldman, M. S. Taqqu, pp. 435-459. Birkhuser,Boston, 1998.
[5] S. McCanne and S. Floyd. ns Network Simulator.http://www.isi.edu/nsnam/ns/
[6] M.E. Gomez and V. Santoja, “Analysis of Self-similarity in I/O Workloadusing Structural Modeling”, �
� �International Symposium on Modeling,
Analysis and Simulation of Computer and Telecommunications Systems,March 24-28, 1999, College Park, Maryland.
[7] D. P Heyman, T. V. Lakshman and A. L. Neidhardt, ”.A New Method forAnalysing Feedback-based Protocols with Applications to EngineeringWeb Traffic over the Internet”, Proc. ACM SIGMETRICS’97 Seattle,USA.
[8] R. Braden, “RFC 1122: Requirements for Internet Hosts – Communica-tion Layers”, October 1989.
[9] 3GPP TS 34.108, “Common Test Environment for User Equipment (UE)Conformance Testing”, version 4.1.0, Release 1999.